{
 "artifacts": null,
 "category": "architecture",
 "description": "Insert a differentiable reaction-diffusion layer that converts dense token or pixel features into sparse, spatially coherent routing masks. Two competing orientations form complexes through conserved monomer reservoirs, so local assignments can cluster while opposite assignments mutually exclude one another instead of independently activating at the same location. The layer can be used as a soft-to-hard MoE router, attention-mask generator, or object-part grouping module.",
 "formulas_latex": [
  "$$\\begin{aligned} A_T\u0026=D_AA_{XX}-kAB^{\\dagger}+vC,\\\\ A^{\\dagger}_T\u0026=D_AA^{\\dagger}_{XX}-kA^{\\dagger}B+vC^{\\dagger},\\\\ B_T\u0026=D_BB_{XX}-kA^{\\dagger}B+vC^{\\dagger},\\\\ B^{\\dagger}_T\u0026=D_BB^{\\dagger}_{XX}-kAB^{\\dagger}+vC,\\\\ C_T\u0026=D_CC_{XX}+kAB^{\\dagger}-vC,\\\\ C^{\\dagger}_T\u0026=D_CC_{XX}+kA^{\\dagger}B-vC^{\\dagger}.\\end{aligned}$$",
  "$$A_X(0,T)=A_X(L,T)=A^{\\dagger}_X(0,T)=A^{\\dagger}_X(L,T)=B_X(0,T)=B_X(L,T)=B^{\\dagger}_X(0,T)=B^{\\dagger}_X(L,T)=C_X(0,T)=C_X(L,T)=C^{\\dagger}_X(0,T)=C^{\\dagger}_X(L,T)=0.$$",
  "$$\\mathbf z^{t+1}=\\mathbf z^t+\\Delta t\\left(\\mathbf D\\,L\\mathbf z^t+\\mathbf r(\\mathbf z^t;\\mathbf h)\\right),\\qquad L\\mathbf 1=0,$$",
  "$$\\begin{aligned}r_C\u0026=kAB^{\\dagger}-vC, \u0026 r_{C^{\\dagger}}\u0026=kA^{\\dagger}B-vC^{\\dagger},\\\\r_A\u0026=-r_C, \u0026 r_{B^{\\dagger}}\u0026=-r_C,\\\\r_{A^{\\dagger}}\u0026=-r_{C^{\\dagger}}, \u0026 r_B\u0026=-r_{C^{\\dagger}}.\\end{aligned}$$"
 ],
 "id": 2847,
 "implementation": "(1) Integration point: place the module after a token, image, or graph encoder and before an MoE router or attention block. Let h_i\\in\\mathbb R^d be the feature at node i, and let the module output p_i=C_i/(C_i+C_i^{\\dagger}+\\epsilon), q_i=C_i+C_i^{\\dagger}, and optionally q_i h_i for routing. Use a 1D sequence Laplacian, 2D four-neighbor image Laplacian, or normalized graph Laplacian. Initialize A_i,B_i,A_i^{\\dagger},B_i^{\\dagger} with softplus projections of h_i plus a positive floor, and initialize C_i=C_i^{\\dagger}=0. (2) Pseudocode: compute k_i=k_0\\exp(g_k(h_i)) and v_i=v_0\\exp(g_v(h_i)); for t=1,...,S compute rC_i=k_i A_i B_i^{\\dagger}-v_i C_i and rCd_i=k_i A_i^{\\dagger}B_i-v_i C_i^{\\dagger}; update each state with z_i\\leftarrow z_i+dt[D_z(Lz)_i+r_{z,i}], using r_A=-rC, r_{B^{\\dagger}}=-rC, r_{A^{\\dagger}}=-rCd, and r_B=-rCd; apply a positivity-preserving clamp or softplus parameterization after each step. For improved conservation, use reaction splitting: perform the local reaction update in several small substeps, then apply conservative graph diffusion. (3) The paper-derived quantities are the reaction stoichiometry, zero-row-sum diffusion, and conserved mass identities. Estimate k_0,v_0,D_A,D_B,D_C and the rate networks by backpropagation. Enforce an explicit-step stability condition dt\\max_z D_z\\rho(-L)\u003c1 and dt\\max_i(k_i\\max(A_i,B_i)+v_i)\u003c1, or use implicit diffusion. Monitor M_{A+C}=\\sum_i(A_i+C_i), M_{B^{\\dagger}+C}=\\sum_i(B_i^{\\dagger}+C_i), M_{A^{\\dagger}+C^{\\dagger}}, and M_{B+C^{\\dagger}}; relative drift should remain below 10^{-5}. (4) First cheap experiment: use CIFAR-10 or MNIST patches with a small CNN or ViT, comparing a standard softmax top-1 router and an unconstrained sigmoid router against this layer at equal parameter count and FLOPs. Measure classification accuracy, assignment entropy, connected-component count, orientation overlap \\sum_i p_i(1-p_i), and robustness to feature noise. A successful result is lower overlap between opposite orientations, more spatially coherent active regions, and equal-or-better accuracy at the same number of routed experts. For language, repeat on a synthetic sequence task where contiguous spans share a latent label; success is higher span consistency without training-loss oscillation.",
 "math_summary": "The paper models two monomers and two oriented complexes on a domain: A and B are monomers for one orientation, A^{\\dagger} and B^{\\dagger} are the corresponding monomers for the opposite orientation, and C and C^{\\dagger} are the two complexes. D_A,D_B,D_C\u003e0 are diffusivities; k\\geq 0 is a local association rate; v\\geq 0 is a local dissociation rate; X is spatial position; T is time. The reaction-diffusion equations are A_T=D_AA_{XX}-kAB^{\\dagger}+vC, A^{\\dagger}_T=D_AA^{\\dagger}_{XX}-kA^{\\dagger}B+vC^{\\dagger}, B_T=D_BB_{XX}-kA^{\\dagger}B+vC^{\\dagger}, B^{\\dagger}_T=D_BB^{\\dagger}_{XX}-kAB^{\\dagger}+vC, C_T=D_CC_{XX}+kAB^{\\dagger}-vC, and C^{\\dagger}_T=D_CC^{\\dagger}_{XX}+kA^{\\dagger}B-vC^{\\dagger}. Zero-flux boundaries impose spatial derivatives equal to zero at the boundary. The reaction terms conserve each participating molecular mass: A+C is conserved by the first reaction, B^{\\dagger}+C is conserved, A^{\\dagger}+C^{\\dagger} is conserved, and B+C^{\\dagger} is conserved, up to diffusion flux. For a neural layer on a finite grid, use a Neumann graph Laplacian L with L\\mathbf{1}=0 and apply explicit updates z^{t+1}=z^t+\\Delta t(DLz^t+r(z^t)). The key structural property is that summing over nodes cancels all reaction terms and gives exact global conservation when L has zero row sum. Rates can be made feature-dependent using positive functions k_i=k_0\\exp(g_k(h_i)) and v_i=v_0\\exp(g_v(h_i)); a saturating feedback variant is k_i=k_0(1+\\alpha C_i/(K+C_i)), while an unbounded variant is k_i=k_0\\exp(\\alpha C_i).",
 "math_tags": [
  "pde",
  "dynamical-systems",
  "optimization"
 ],
 "ml_areas": [
  "moe-routing",
  "attention",
  "graph-nn"
 ],
 "paper": {
  "arxiv_id": "2608.29679",
  "arxiv_url": "https://arxiv.org/abs/2608.29679",
  "summary_what_math_gives_to_ml": "This paper provides a constructive mass-conserving reaction-diffusion mechanism that separates three useful behaviors: total punctum mass, punctum width, and punctum spacing. Its transferable asset is not merely diffusion, but reversible local reactions with conserved reservoirs and sign-opposed orientation channels, which can create sparse, mutually exclusive assignments without softmax normalization. A neural implementation could use a small number of explicit reaction-diffusion steps as a differentiable clustering or routing layer over image patches, tokens, or graph nodes, with feature-dependent reaction rates and exact conservation enforced by the update. The reported absence of oscillatory instabilities also makes this a plausible stable alternative to unconstrained recurrent competitive routing.",
  "title": "Clustering versus sorting: a mass-conserving reaction-diffusion model of planar polarity puncta",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 5,
  "novelty": 7,
  "usefulness": 6
 },
 "solves": [
  "accuracy",
  "stability",
  "sample-efficiency"
 ],
 "title": "Mass-Conserving Puncta Router",
 "url": "https://synthcore.org/idea/2847/mass-conserving-puncta-router",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}